The rumors had it right: Sony is rebooting its robot dog, Aibo, announcing a new four-legged companion AI-powered bot incoming with the same brand name but more rounded corners and visible facial features for extra kawaii, including a pair of expressive, puppy-dog eyes. Deep learning tech, fish-eye cameras and a series of other embedded sensors enable Aibo to detect and analyze sounds and images so that it can learn and respond to its environment and interact with its owner so it appears less, well, robotic. Sony claims Aibo's adaptive behavior includes being able to actively seek out its owners; detect words of praise; smiles; head and back scratches; petting, and more. Thanks to the embedded cameras you can also instruct Aibo to take a photo for you -- should you want a dog's eye view of yourself/your home life. "Aibo's AI learns from interactions with its owners and develops a unique personality over time," it writes.

Deep learning promises to make computers far better at tasks like facial recognition, image understanding and machine translation. Our October puzzle asked you to play with simple artificial neural networks to explore the phenomenon of deep learning and gain some insight into how it has achieved such spectacular successes in the field of artificial intelligence. We're going to create a simple network that converts binary numbers to decimal numbers. Imagine a network with just two layers: an input layer consisting of three units and an output layer with seven units. Each unit in the first layer connects to each unit in the second, as shown in the figure below.

Geoffrey Hinton may be the "godfather" of deep learning, a suddenly hot field of artificial intelligence, or AI – but that doesn't mean he's resting on his algorithms. Hinton, a University Professor Emeritus at the University of Toronto, recently released two new papers that promise to improve the way machines understand the world through images or video – a technology with applications ranging from self-driving cars to making medical diagnoses. "This is a much more robust way to detect objects than what we have at present," Hinton, who is also a fellow at Google's AI research arm, said today at a tech conference in Toronto. "If you've been in the field for a long time like I have, you know that the neural nets that we use now – there's nothing special about them. We just sort of made them up."

We're working with Booz Allen Hamilton to help the U.S. federal government apply deep learning techniques to key challenges in healthcare, defense and cybersecurity. Certified Deep Learning Institute instructors from NVIDIA and Booz Allen will provide hands-on training to federal customers across a variety of government agencies to build deep learning and data-driven solutions that are needed in the field. "The Deep Learning Institute has developed industry-leading curricula with the world's leading AI experts, and we deliver that in hands-on classes taught by certified instructors," said Greg Estes, vice president of Developer Programs at NVIDIA. "By working together with Booz Allen Hamilton, we will train specialists and data scientists to help tackle complex challenges that confront the federal government in healthcare, cybersecurity and other important areas." "Deep learning and AI-first approaches are critical to every federal agency. Booz Allen and NVIDIA working together to meet the demand of the federal government for training and applying deep learning techniques will further innovation," said Dr. Josh Sullivan, a senior vice president who leads Booz Allen's data science capabilities.

Picture the hospital of the future replete with a NASA-like command center featuring scores of information screens and a radiology department that leverages AI technology to help improve diagnostic accuracy and deep-learning technology to ensure that radiology images are clear. This is the world of data-driven medicine that, sees -- not in a crystal ball but in the real world. "Those technologies are here now, and they are gaining steam," said Charles Koontz, CEO of GE Healthcare Digital and chief digital officer of GE Healthcare in an interview at GE Digital's Minds Machines event in San Francisco last week. A 2016 McKinsey study supports the notion that the healthcare sector is embracing digital transformation. The field has seen "some core change," according to McKinsey, basing that assessment on a survey of 10 verticals.

Machine learning is usually defined in behaviourist terms, where external validation is the primary mechanism of learning. In this paper, I argue for a more holistic interpretation in which finding more probable, efficient and abstract representations is as central to learning as performance. In other words, machine learning should be extended with strategies to reason over its own learning process, leading to so-called meta-cognitive machine learning. As such, the de facto definition of machine learning should be reformulated in these intrinsically multi-objective terms, taking into account not only the task performance but also internal learning objectives. To this end, we suggest a "model entropy function" to be defined that quantifies the efficiency of the internal learning processes. It is conjured that the minimization of this model entropy leads to concept formation. Besides philosophical aspects, some initial illustrations are included to support the claims.

Many real world tasks such as reasoning and physical interaction require identification and manipulation of conceptual entities. A first step towards solving these tasks is the automated discovery of distributed symbol-like representations. In this paper, we explicitly formalize this problem as inference in a spatial mixture model where each component is parametrized by a neural network. Based on the Expectation Maximization framework we then derive a differentiable clustering method that simultaneously learns how to group and represent individual entities. We evaluate our method on the (sequential) perceptual grouping task and find that it is able to accurately recover the constituent objects. We demonstrate that the learned representations are useful for next-step prediction.

In the last years deep learning has demonstrated a great success in a large number of fields and has attracted the attention of various research communities with the consequent development of multiple coding frameworks (e.g., Caffe [Jia et al., 2014], TensorFlow [Abadi et al., 2015]), the diffusion of blogs, online tutorials, books, and dedicated courses. Besides reaching out scientists with different backgrounds, the need of all these supportive tools originates also from the nature of deep learning: it is a methodology that involves many structural details as well as several hyperparameters whose importance has been growing with the recent trend of designing deeper and multi-branches networks. Some of the hyperparameters define the model itself (e.g., number of hidden layers, regularization coefficients, kernel size for convolutional layers), while others are related to the model training procedure. In both cases, hyperparameter tuning is a critical step to realize deep learning full potential and most of the knowledge in this area comes from living practice, years of experimentation, and, to some extent, mathematical justification [Bengio, 2012]. With respect to the optimization process, stochastic gradient descent (SGD) has proved itself to be a key component of the deep learning success, but its effectiveness strictly depends on the choice of the initial learning rate and learning rate schedule. This has primed a line of research on algorithms to reduce the hyperparameter dependence in SGD--see Section 2 for an overview on the related literature.

From only positive (P) and unlabeled (U) data, a binary classifier could be trained with PU learning, in which the state of the art is unbiased PU learning. However, if its model is very flexible, empirical risks on training data will go negative, and we will suffer from serious overfitting. In this paper, we propose a non-negative risk estimator for PU learning: when getting minimized, it is more robust against overfitting, and thus we are able to use very flexible models (such as deep neural networks) given limited P data. Moreover, we analyze the bias, consistency, and mean-squared-error reduction of the proposed risk estimator, and bound the estimation error of the resulting empirical risk minimizer. Experiments demonstrate that our risk estimator fixes the overfitting problem of its unbiased counterparts.

As developers flock to artificial intelligence frameworks in response to the explosion of intelligence machines, training deep learning models has emerged as a priority along with synching them to a growing list of neural and other network designs. All are being aligned to confront some of the next big AI challenges, including training deep learning models to make inferences from the fire hose of unstructured data. These and other AI developer challenges were highlighted during this week's Nvidia GPU technology conference in Washington. The GPU leader uses the events to bolster its contention that GPUs--some with up to 5,000 cores--are filling the computing gap created by the decline of Moore's Law. The other driving force behind the "era of AI" is the emergence of algorithm-driven deep learning that is forcing developers to move beyond mere coding to apply AI to a growing range of automated processes and predictive analytics.